Enhancing learning algorithms by an effective structure-based dissimilarity measuring approach

Many unsupervised and supervised learning algorithms are based on how well measuring the dissimilarity between objects is performed. Existing dissimilarity measures such as Euclidean and dynamic time warping distances often treat all measurements describing an object equally. This consideration might not be appropriate for the objects described by features at two different levels of abstraction. In order to distinguish such objects from each other and enhance distance-based learning algorithms, we propose a simple but effective structure-based dissimilarity measuring approach. The proposed approach is additive to enable the key features at the first level to contribute to the measure equally and independently. For the second level, it preserves the properties of any employed measure to compute the dissimilarity between the objects based on the detailed components of each key feature. Experiments on real and standardized data sets show that our approach can make several popular distance-based learning algorithms more effective.